Hi, joe The fundamental correctness is guaranteed by the fencing mechanism provided by Apache BookKeeper and the CAS operation provided by the metadata storage. Both fencing and CAS operations will prevent two owners updating data or metadata at the same time. This default shutdown behavior isn’t changed. We just introduce an alternative way for improving stability when Zookeeper is doing leader election or becomes unavailable.
According to the following rules, I think this will not break current system principles. 1. If the znode for the bundle is deleted, this is consistent with the current behavior. The broker who acquires for the lock first will become the owner. 2. If the znode for the bundle is not deleted, other brokers also unable to acquire the lock. Both the broker that re-creates the session and other brokers need to wait for the znode deletion. Then the broker who acquires for the lock first will become the owner. 3. If the bundle ownership changed, the broker that re-creates the session unable to acquire the lock. So the broker should unload the bundle, This is also consistent with current ownership change behavior. 4. Also, if unexpected exceptions throw during the re-own process, the broker needs to shutdown itself. Thanks, Penghui On Feb 22, 2020, 12:27 +0800, PengHui Li <codelipeng...@gmail.com>, wrote: > Hi all, > > I have drafted a proposal for improving broker's Zookeeper session timeout > handling. You can find at > https://github.com/apache/pulsar/wiki/PIP-57%3A-Improve-Broker%27s-Zookeeper-Session-Timeout-Handling > > Also I copy it to the email thread for easier to view. Any suggestions or > ideas welcome to join the discussion. > > > PIP 57: Improve Broker's Zookeeper Session Timeout Handling > Motivation > In Pulsar, brokers use Zookeeper as the configuration store and broker > metadata maintaining. We can also call them Global Zookeeper and Local > Zookeeper. > The Global Zookeeper maintains the namespace policies, cluster metadata, and > partitioned topic metadata. To reduce read operations on Zookeeper, each > broker has a cache for global Zookeeper. The Global Zookeeper cache updates > on znode changed. Currently, when the present session timeout happens on > global Zookeeper, a new session starts. Broker does not create any EPHEMERAL > znodes on global Zookeeper. > The Local Zookeeper maintains the local cluster metadata, such as broker load > data, topic ownership data, managed ledger metadata, and Bookie rack > information. All of broker load data and topic ownership data are create > EPHEMERAL nodes on Local Zookeeper. Currently, when session timeout happens > on Local Zookeeper, the broker shutdown itself. > Shutdown broker results in ownership change of topics that the broker owned. > However, we encountered lots of problems related to the current session > timeout handling. Such as broker with long JVM GC pause, Local Zookeeper > under high workload. Especially the latter may cause all broker shutdowns. > So, the purpose of this proposal is to improve session timeout handling on > Local Zookeeper to avoid unnecessary broker shutdown. > Approach > Same as the Global Zookeeper session timeout handling and Zookeeper session > timeout handling in BookKeeper, a new session should start when the present > session timeout. > If a new session failed to start, the broker would retry several times. The > retry times depend on the configuration of the broker. After the number of > retries, if still can't start session success, the broker still needs to be > shut down since this may be a problem with the Zookeeper cluster. The user > needs to restart the broker after the zookeeper cluster returns to normal. > If a new session starts success, the issue is slightly more complicated. So, > I will introduce every scene separately. > Topic ownership data handling > The topic ownership data maintain all namespace bundles that owned by the > broker. In Zookeeper, create an EPHEMERAL znode for each namespace bundle. > When the session timeout happens on the local Zookeeper, all of the EPHEMERAL > znode maintained by this broker will delete automatically. We need some > mechanism to avoid the unnecessary ownership transfer of the bundles. Since > the broker cached the owned bundles in memory, the broker can use the cache > to re-own the bundles. > Firstly, when the broker to re-own the bundle, if the znode of the bundle > exists at Zookeeper and the owner is this broker, it may be that Zookeeper > has not deleted the znode. The broker should check if the ephemeral owner is > the current session ID. If not, the broker should wait for the znode deletion. > Then the broker tries to own the bundle. If the broker owns the bundle > success means the bundle not owned by other brokers, the broker should check > whether to preload the topics under the bundle. If the broker failed to own > the bundle means the bundle owned by another broker. The broker should unload > the bundle. > Theoretically, the mechanism can guarantee that the ownership of most bundles > will not change during the session timeout. > Broker load data handling > The load data used for namespace bundle load balancing, so there is no need > to be overly complicated in handling. The only effect is that it will > interfere with the choice of the broker when finding a candidate broker for a > namespace bundle. Even without selecting the optimal broker, it will continue > to relocate the namespace bundles. > So for broker load data handling, we need to guarantee the load data of the > broker can report success. > Other scene handing > There are also some usage scenarios of the local Zookeeper, BookKeeper > client, managed ledger meta, bookie rack information, and schema metadata. > All of these scenarios do not create any EPHEMERAL znodes on the Zookeeper. > Pulsar introduces the Zookeeper cache for the local Zookeeper. The cache is > invalidated when the session timeout occurs. > Configurations > A new configuration parameter zookeeperSessionExpiredPolicy added to > broker.conf to control the zookeeper session expired policy. There are two > options, shutdown and reconnect. > > > Thanks, > Penghui
